My course projects at IMPA can be accesed from here.

Fall 2013

MAX2CORR in cubic graphs

Course: Approximation Algorithms
Professor: Vladimir Braverman

Abstract:We present a probably good approximation of MAX2CORR in cubic graphs. This is obtained by adapting Zwicks[8] CSDP+LI approach of MAXCUT in cubic graphs. The new factor outperforms the 1/3 approximation factor obtained independently from the Greedy and the SDP approach.


Spring 2014

Emotion Recognition using joint Low Rank and Sparse Representation

Course: Compressed Sensing and Sparse Representation
Professor: Truc D. Tran

Abstract:We present an approach to emotion recognition applying a join sparse and low rank decomposition. Previous sparse-based approaches to this problem require of explicitly provided neutral faces to assist the recognition task. Our model trade the neutral face requirement by a sound prior: neutral face can be capture as the low rank component of frame sequence with moderate variation. Satisfactory recognition rates were attained on 7 different emotions.


Design and Implementation of Constrained-RPCA Decomposition in CUDA

Course: Parallel Programming
Professor: Randall Burns

Abstract: In this project we explore the convenience of massively parallel processors architectures for the design and implementation of the Alternating Direction Method of Multipliers(ADMM) on the solution of a convex optimization problem that decompose a signal into a low rank component and a sparse error.


Fall 2014

Tool maipulation through Leap Motion

Course: Medical Augmented Reality
Professor: Nassir Navab

Abstract:In this project we explore visual clues that improve depth perception and facilitate the manipulation of a tool in a virtual environment. We use Leap Motion for tool tracking. We consider shadows, mirror reflections, heat maps, and reference objects as our visual clues.


Spring 2016

Low Dimensional Embedding of a Pose Collection

Course: Modeling and segmentation of high dimensional data
Professor: Rene Vidal

Abstract: In this project we evaluate how the problem of identifying exemplars within a collection (in the form K pairwise most-dissimilar elements) can be simplified by reducing the input set to the convex hull of a low dimensional embedding. We use Laplacian Eigenmaps(LE) and Muli-Dimensional Scaling(MDS) as embedding techniques. Computation of convex hull in low dimensions (d=2,3) is very efficient (O(n logn)). In our test example the convex hull provide an efficient subset selection (1%-3% of the original collection) while preserving the most dissimilar poses.


Detection of microcalcifications

Course: Medical Image Analysis
Professor: Jerry L. Prince

Abstract: We study the problem of detecting microcalcifications on mammograms. We segment the breast using intensity based techniques and get a more precise boundary using active contours. We grow candidate microcalcifications regions and classify them according to size, shape and concentration. Our classification error was 60% in images with microcalcifications and 80% in normal images.


Thalamus segmentation through brain registration

Course: Medical Image Analysis
Professor: Jerry L. Prince

Abstract: We propouse a fast brain registration technique based on morphological symmetry. For each brain model we identify the sagital plane by computing the principal direction of the brain mask and identifying the symmetry axis on the coronal view. We compute registration of pairs of brains on the space of rigid transformation following an alternating optimization approach.